Neural Network Driven Inference of One-Carbon Metabolic Flux from Isotope Labeling

Abstract

Metabolism is a network of interconnected pathways that provide cells with the energy needed to function and sustain life. In mammalian tissues, metabolites are processed through thousands of reactions to fulfill different metabolic tasks. Dysregulation of these metabolic tasks causes disruptions in metabolic fluxes, which has been connected to many leading causes of death, including cancer, diabetes, and cardiovascular diseases. Stable isotope tracing, which labels molecules, and mass spectrometry, which measures isotopologue distributions from labeled metabolites, have played an important role in studying these metabolic tasks. Metabolic Flux Analysis (MFA) can be applied to infer metabolic fluxes from labeling observations. Determining these fluxes is difficult because the relationships between fluxes and isotopologue distributions are nonlinear and information from different pathways often overlap for a single isotopologue measurement. Even in small-scale systems, understanding how isotope labeling relates to metabolic flux has been unclear. Here we show that machine learning can accurately predict simulated metabolic fluxes from isotope labeling data. Using the serine–glycine one-carbon system as a case study, we demonstrate that a full set of fluxes can be predicted using only two tracers. Within this framework, a neural network infers flux more robustly than a standard linear regression model. These findings provide a proof of concept for neural network driven metabolic flux inference. The framework developed on this thesis will serve as a foundation to identify flux using machine learning in more complex systems. Specifically, future work will apply this framework to build an atlas of mammalian metabolism, such as a flux map for major organs.